search

ozendelait / rvc_devkit

Robust Vision Challenge Devkits
http://www.robustvision.net/
MIT License
107 stars 13 forks source link

A Classes Number Question #47

Closed LLccelerator closed 2 years ago

LLccelerator commented 2 years ago

We are preparing for Robust Vision Challenge 2022(Obj. Det.), but we find a question recently.As you asked, our label space shoud not exceed 550 classes, while the OID dataset has 600 classes. So We do not think this limit is very reasonable.Is there any solutions or can you give us some other advice? Thanks!

ozendelait commented 2 years ago

OID boxable should have 500 classes as a maximum (see https://www.kaggle.com/competitions/open-images-object-detection-rvc-2022-edition and https://github.com/ozendelait/rvc_devkit/blob/master/common/label_definitions/oid-challenge-2019-classes-description-500.csv ); I have seen the number 600 pop up on the OID website. @akuznetso : please comment

akuznetso commented 2 years ago

We do think the subset of 500 classes we selected for the challenge is reasonable. We removed some very broad classes (e.g. "clothing") and some infrequent ones (e.g. "paper cutter"). This is to enable more precise mAP measurements.

ozendelait commented 2 years ago

Thanks Alina for the clarification. @LLccelerator: 550 classes should be enough given the overlaps of classes between OID, COCO, MVD We can accept your solution if you overshoot by 10 classes given you describe where you think our assumed overlap was unfounded.